A CTO asking “are we compliant with the AI Act’s marking rules” is asking the wrong question first. Article 50 tells you that AI-generated content needs to be machine-readably marked, effective, interoperable, robust and reliable, as far as technically feasible. It does not tell you which specification to implement, how to sign a manifest, or what to do when a CDN strips your metadata on the way to a browser. Meanwhile, Content Credentials built on the C2PA specification are already running at platform scale on Google Pixel cameras, inside YouTube uploads, across TikTok’s video library and on LinkedIn’s feed, independent of any single jurisdiction’s law.
The practical argument for engineering teams is this: build to C2PA’s Content Credentials architecture, because it is the standard with real production deployments, real tooling, and a real (if incomplete) verification record, rather than architecting primarily around a regulatory clause that names an outcome without naming a mechanism. The Article 50 obligation and the C2PA specification point at the same technical shape, a signed manifest paired with a robustness layer, but only one of them ships code, SDKs and a conformance programme you can build against today.
What the regulation asks for, and what it deliberately leaves open
Article 50(2) requires marking to be effective, interoperable, robust and reliable “as far as technically feasible”, and the EU AI Act’s Code of Practice adds that at least two machine-readable marking layers are needed for most modalities, because no single technique currently meets all four criteria at once. That is a fair technical judgement rather than a drafting failure: it is consistent with what the standards themselves report about their own limits. But it leaves the actual engineering decisions, which specification, which signing infrastructure, which fallback layer, entirely to the team building the pipeline. (We have covered the compliance mechanics of that gap, including where provenance dies inside a typical delivery chain, in a companion explainer on machine-readable marking.)
That gap is where C2PA has moved fastest. The specification did not wait for a regulator to finish drafting a Code of Practice; it shipped v2.3 early in 2026 and v2.4 in April 2026, with the Content Authenticity Initiative maintaining official SDKs across Rust, JavaScript, Python, C++, Swift and Android. Regulation describes an obligation. C2PA describes a manifest format, a signing process and a conformance programme, and those are the things an engineering team actually implements.
What is actually running in production
The scale of adoption is the part most compliance-first framing misses. Adobe’s Photoshop, Lightroom and Firefly, Microsoft’s Bing Image Creator and M365, and device makers including Leica, Nikon, Canon and Samsung have implemented or announced C2PA support. Google says its Pixel 10 is the first device to reach Assurance Level 2, the highest level in the C2PA Conformance Program, although Samsung’s Galaxy S25 shipped C2PA support earlier. Google Search’s “About this image” feature reads C2PA data, YouTube surfaces Content Credentials on uploads, TikTok has labelled more than 3 billion videos, using Content Credentials alongside its own watermarks and labels, and LinkedIn preserves and displays Content Credentials in-feed. France Télévisions became the first broadcaster to sign its France 2 news programme with C2PA systematically, every day. The Content Authenticity Initiative reported more than 6,000 member organisations in January 2026.
Regulation was only one of the drivers. The 2025-2026 acceleration in provenance adoption traces to three converging pressures, election-misinformation concerns, EU AI Act provisions and platform-policy requirements, rather than to any single regulatory mandate. Platforms also had reasons of their own: content provenance became a user-facing trust signal as well as a compliance artefact, and the specification already had multi-language tooling that a team could integrate rather than build from a legal text. An industry-analyst estimate from SNS Insider projects the AI watermarking market to exceed $5,040.94 million by 2035, driven by authentication and copyright-protection demand rather than any single jurisdiction’s rulebook; treat that as a market-research projection, not a settled fact, but it is directionally consistent with what the adoption data above already shows.
The part production adoption does not cover: independent verification
Building to what is deployed is not the same as building to what has been independently proven safe, and the C2PA record makes that distinction concrete. The specification reached device- and platform-scale adoption roughly two years before it received its first comprehensive, independent security analysis. That analysis, the first formal-methods review of the C2PA specification, was published in April 2026. Its authors found that the specifications “fail to achieve their claimed security goals” and recommended against relying on them in high-stakes applications such as financial disclosure, journalism or legal evidence until the issues it identifies are fixed. It arrived after Pixel 10 certification, after TikTok’s large-scale labelling, after most of the mainstream adoption described above had already happened.
That does not mean avoiding C2PA. It means treating “is this standard widely adopted” and “has this standard been independently verified” as two separate engineering questions, because the gap between them is exactly where risk concentrates for anything you plan to rely on as a control rather than a label. Truepic, a C2PA founding member, sells an “Enterprise C2PA” layer that adds security and support on top of the open specification. The maintainers’ own c2pa-attacks tooling exists because the standard’s stewards treat robustness against manipulation as an open, ongoing problem, not a solved one.
What to actually implement
For a team building a content pipeline that needs to survive both regulatory scrutiny and an adversarial internet, four decisions follow directly from the evidence above.
Sign manifests at the point of generation, not afterwards. A C2PA manifest is a cryptographically signed set of assertions about origin, edit history and AI involvement. It has to be created where the content is created; bolting it on later loses the chain-of-custody guarantee that makes it worth having.
Decide your trust list before you decide your signing certificate. Validating a manifest depends on the verifier accepting the certificate authority that issued it. C2PA maintains an official Trust List, but which certificate authorities your verification pipeline accepts is still a governance decision your team has to make explicitly.
Do not rely on the manifest alone to survive transformation. C2PA’s own mechanism for surviving file rewrites, soft binding, delegates robustness to an embedded watermark rather than solving it in the manifest itself. Any pipeline that resizes, re-encodes or transcodes content between generation and publication (which is most of them) needs a second layer, consistent with the Code of Practice’s two-layer requirement and with C2PA’s own architecture.
Build cross-layer verification, because no vendor sells it yet. Work accepted at the CVPR 2026 APAI workshop demonstrated an “integrity clash”: a cryptographically valid C2PA manifest asserting human authorship over content that separately carries an AI watermark, with each layer passing verification independently because nothing checks them against each other. Their cross-layer verification protocol reported 100% classification accuracy across 3,500 test images, and the authors describe closing the gap between the two layers as technically straightforward. No commercially available product in this research shipped that check, which makes it a build problem for the engineering team rather than a shopping list item.
Score any provenance stack you inherit or evaluate on three axes independently: how widely adopted it is, how independently verifiable it is, and how it holds up under real transformation, not just under vendor benchmarks. No standard identified in this research scores well on all three at once, which argues for a due-diligence process rather than waiting for a winner to emerge. Article 50 will keep evolving its language. The manifests, signing infrastructure and cross-layer checks a team builds against C2PA’s Content Credentials architecture today are work that survives the next redraft, because they are engineering decisions rather than compliance interpretations.
Zartis’s AI Governance practice runs this due-diligence process for clients evaluating or building a content-provenance stack, and our AI Development teams build the manifest signing, trust-list configuration and cross-layer verification the standards themselves do not ship. Talk to us about a provenance and content-credentials assessment.
